Available online at www.sciencedirect.
com
Reinforcement learning: The Good, The Bad and The Ugly
Peter Dayana and Yael Nivb
Reinforcement learning provides both qualitative and
quantitative frameworks for understanding and modeling
adaptive decision-making in the face of rewards and
punishments. Here we review the latest dispatches from the
forefront of this field, and map out some of the territories where
lie monsters.
Addresses
a
UCL, United Kingdom
b
Psychology Department and Princeton Neuroscience Institute,
Princeton University, United States
Corresponding authors: Dayan, Peter (dayan@gatsby.ucl.ac.uk) and
Niv, Yael (yael@princeton.edu)
Current Opinion in Neurobiology 2008, 18:185–196
This review comes from a themed issue on
Cognitive neuroscience
Edited by Read Montague and John Assad
Available online 22nd August 2008
0959-4388/$ – see front matter
# 2008 Elsevier Ltd. All rights reserved.
DOI 10.1016/j.conb.2008.08.003
Introduction
Reinforcement learning (RL) [1] studies the way that
natural and artificial systems can learn to predict the con-
sequences of and optimize their behavior in environments
in which actions lead them from one state or situation to the
next, and can also lead to rewards and punishments. Such
environments arise in a wide range of fields, including
ethology, economics, psychology, and control theory.
Animals, from the most humble to the most immodest,
face a range of such optimization problems [2], and, to an
apparently impressive extent, solve them effectively. RL,
originally born out of mathematical psychology and oper-
ations research, provides qualitative and quantitative com-
putational-level models of these solutions.
However, the reason for this review is the increasing
realization that RL may offer more than just a compu-
tational, ‘approximate ideal learner’ theory for affective
decision-making. RL algorithms, such as the temporal
difference (TD) learning rule [3], appear to be directly
instantiated in neural mechanisms, such as the phasic
activity of dopamine neurons [4]. That RL appears to be
so transparently embedded has made it possible to use it
in a much more immediate way to make hypotheses
about, and retrodictive and predictive interpretations
of, a wealth of behavioral and neural data collected in a
www.sciencedirect.com
huge range of paradigms and systems. The literature in
this area is by now extensive, and has been the topic of
many recent reviews (including [5–9]). This is in addition
to rapidly accumulating literature on the partly related
questions of optimal decision-making in situations invol-
ving slowly amounting information, or social factors such
as games [10–12]. Thus here, after providing a brief
sketch of the overall RL scheme for control (for a more
extensive review, see [13]), we focus only on some of the
many latest results relevant to RL and its neural instan-
tiation. We categorize these recent findings into those
that fit comfortably with, or flesh out, accepted notions
(playfully, ‘The Good’), some new findings that are not as
snugly accommodated, but suggest the need for exten-
sions or modifications (‘The Bad’), and finally some key
areas whose relative neglect by the field is threatening to
impede its further progress (‘The Ugly’).
The reinforcement learning framework
Decision-making environments are characterized by a
few key concepts: a state space (states are such things
as locations in a maze, the existence or absence of
different stimuli in an operant box or board positions
in a game), a set of actions (directions of travel, presses on
different levers, and moves on a board), and affectively
important outcomes (finding cheese, obtaining water, and
winning). Actions can move the decision-maker from one
state to another (i.e. induce state transitions) and they can
produce outcomes. The outcomes are assumed to have
numerical (positive or negative) utilities, which can
change according to the motivational state of the
decision-maker (e.g. food is less valuable to a satiated
animal) or direct experimental manipulation (e.g. poison-
ing). Typically, the decision-maker starts off not knowing
the rules of the environment (the transitions and out-
comes engendered by the actions), and has to learn or
sample these from experience.
In instrumental conditioning, animals learn to choose
actions to obtain rewards and avoid punishments, or,
more generally to achieve goals. Various goals are
possible, such as optimizing the average rate of acqui-
sition of net rewards (i.e. rewards minus punishments), or
some proxy for this such as the expected sum of future
rewards, where outcomes received in the far future are
discounted compared with outcomes received more
immediately. It is the long-term nature of these goals
that necessitates consideration not only of immediate
outcomes but also of state transitions, and makes choice
interesting and difficult. In terms of RL, instrumental
conditioning concerns optimal choice, that is determining
Current Opinion in Neurobiology 2008, 18:185–196
186 Cognitive neuroscience
an assignment of actions to states, also known as a policy
that optimizes the subject’s goals.
By contrast with instrumental conditioning, Pavlovian (or
classical) conditioning traditionally treats how subjects
learn to predict their fate in those cases in which they
cannot actually influence it. Indeed, although RL is
primarily concerned with situations in which action selec-
tion is germane, such predictions play a major role in
assessing the effects of different actions, and thereby in
optimizing policies. In Pavlovian conditioning, the pre-
dictions also lead to responses. However, unlike the
flexible policies that are learned via instrumental con-
ditioning, Pavlovian responses are hard-wired to the
nature and emotional valence of the outcomes.
RL methods can be divided into two broad classes, model-
based and model-free, which perform optimization in very
different ways (Box 1, [14]). Model-based RL uses experi-
ence to construct an internal model of the transitions and
immediate outcomes in the environment. Appropriate
actions are then chosen by searching or planning in this
world model. This is a statistically efficient way to use
experience, as each morsel of information from the
environment can be stored in a statistically faithful and
computationally manipulable way. Provided that constant
replanning is possible, this allows action selection to be
readily adaptive to changes in the transition contingencies
and the utilities of the outcomes. This flexibility makes
model-based RL suitable for supporting goal-directed
actions, in the terms of Dickinson and Balleine [15].
For instance, in model-based RL, performance of actions
leading to rewards whose utilities have decreased is
immediately diminished. Via this identification and other
findings, the behavioral neuroscience of such goal-
directed actions suggests a key role in model-based RL
(or at least in its components such as outcome evaluation)
for the dorsomedial striatum (or its primate homologue,
the caudate nucleus), prelimbic prefrontal cortex, the
orbitofrontal cortex, the medial prefrontal cortex,1 and
parts of the amygdala [9,17–20].
Model-free RL, on the other hand, uses experience to
learn directly one or both of two simpler quantities (state/
action values or policies) which can achieve the same
optimal behavior but without estimation or use of a world
model. Given a policy, a state has a value, defined in terms
of the future utility that is expected to accrue starting
from that state. Crucially, correct values satisfy a set of
mutual consistency conditions: a state can have a high
value only if the actions specified by the policy at that
state lead to immediate outcomes with high utilities, and/
or states which promise large future expected utilities (i.e.
1
Note that here and henceforth we lump together results from rat and
primate prefrontal cortical areas for the sake of brevity and simplicity,
despite important and contentious differences [16].
Current Opinion in Neurobiology 2008, 18:185–196
have high values themselves). Model-free learning rules
such as the temporal difference (TD) rule [3] define any
momentary inconsistency as a prediction error, and use it to
specify rules for plasticity that allow learning of more
accurate values and decreased inconsistencies. Given
correct values, it is possible to improve the policy by
preferring those actions that lead to higher utility out-
comes and higher valued states. Direct model-free
methods for improving policies without even acquiring
values are also known [21].
Model-free methods are statistically less efficient than
model-based methods, because information from the
environment is combined with previous, and possibly
erroneous, estimates or beliefs about state values, rather
than being used directly. The information is also stored
in scalar quantities from which specific knowledge about
rewards or transitions cannot later be disentangled. As
a result, these methods cannot adapt appropriately
quickly to changes in contingency and outcome utilities.
Based on the latter characteristic, model-free RL has
been suggested as a model of habitual actions [14,15], in
which areas such as the dorsolateral striatum and the
amygdala are believed to play a key role [17,18]. How-
ever, a far more direct link between model-free RL and
the workings of affective decision-making is apparent in
the findings that the phasic activity of dopamine
neurons during appetitive conditioning (and indeed
the fMRI BOLD signal in the ventral striatum of
humans, a key target of dopamine projections) has many
of the quantitative characteristics of the TD prediction
error that is key to learning model-free values [4,22–24].
It is these latter results that underpin the bulk of neural
RL.
‘The Good’: new findings in neural RL
Daw et al. [5] sketched a framework very similar to this,
and reviewed the then current literature which pertained
to it. Our first goal is to update this analysis of the
literature. In particular, courtesy of a wealth of exper-
iments, just two years later we now know much more
about the functional organization of RL systems in the
brain, the pathways influencing the computation of pre-
diction errors, and time-discounting. A substantial frac-
tion of this work involves mapping the extensive findings
from rodents and primates onto the human brain, largely
using innovative experimental designs while measuring
the fMRI BOLD signal. This research has proven most
fruitful, especially in terms of tracking prediction error
signals in the human brain. However, in considering these
results it is important to remember that the BOLD signal
is a measure of oxyhemoglobin levels and not neural
activity, let alone dopamine. That neuromodulators can
act directly on capillary dilation, and that even the limited
evidence we have about the coupling between synaptic
drive or neural activity and the BOLD signal is confined
to the cortex (e.g. [25]) rather than the striatum or mid-
www.sciencedirect.com
 Reinforcement learning: The Good, The Bad and The Ugly Dayan and Niv 187

Box 1 Model-based and model-free reinforcement learning

Reinforcement learning methods can broadly be divided into two classes, model-based and model-free. Consider the problem illustrated in the
figure, of deciding which route to take on the way home from work on Friday evening. We can abstract this task as having states (in this case,
locations, notably of junctions), actions (e.g. going straight on or turning left or right at every intersection), probabilities of transitioning from one
state to another when a certain action is taken (these transitions are not necessarily deterministic, e.g. due to road works and bypasses), and
positive or negative outcomes (i.e. rewards or costs) at each transition from scenery, traffic jams, fuel consumed, etc. (which are again
probabilistic).
Model-based computation, illustrated in the left ‘thought bubble’, is akin to searching a mental map (a forward model of the task) that has been
learned based on previous experience. This forward model comprises knowledge of the characteristics of the task, notably, the probabilities of
different transitions and different immediate outcomes. Model-based action selection proceeds by searching the mental map to work out the long-
run value of each action at the current state in terms of the expected reward of the whole route home, and chooses the action that has the highest
value.
Model-free action selection, by contrast, is based on learning these long-run values of actions (or a preference order between actions) without
either building or searching through a model. RL provides a number of methods for doing this, in which learning is based on momentary
inconsistencies between successive estimates of these values along sample trajectories. These values, sometimes called cached values because
of the way they store experience, encompass all future probabilistic transitions and rewards in a single scalar number that denotes the overall future
worth of an action (or its attractiveness compared with other actions). For instance, as illustrated in the right ‘thought bubble’, experience may have
taught the commuter that on Friday evenings the best action at this intersection is to continue straight and avoid the freeway.
Model-free methods are clearly easier to use in terms of online decision-making; however, much trial-and-error experience is required to make the
values be good estimates of future consequences. Moreover, the cached values are inherently inflexible: although hearing about an unexpected
traffic jam on the radio can immediately affect action selection that is based on a forward model, the effect of the traffic jam on a cached propensity
such as ‘avoid the freeway on Friday evening’ cannot be calculated without further trial-and-error learning on days in which this traffic jam occurs.
Changes in the goal of behavior, as when moving to a new house, also expose the differences between the methods: whereas model-based
decision making can be immediately sensitive to such a goal-shift, cached values are again slow to change appropriately. Indeed, many of us have
experienced this directly in daily life after moving house. We clearly know the location of our new home, and can make our way to it by
concentrating on the new route; but we can occasionally take an habitual wrong turn toward the old address if our minds wander. Such
introspection, and a wealth of rigorous behavioral studies (see [15], for a review) suggests that the brain employs both model-free and model-based
decision-making strategies in parallel, with each dominating in different circumstances [14]. Indeed, somewhat different neural substrates underlie
each one [17].

www.sciencedirect.com Current Opinion in Neurobiology 2008, 18:185–196

188 Cognitive neuroscience
brain areas such as the ventral tegmental area, imply that
this fMRI evidence is alas very indirect.
Functional organization: in terms of the mapping from
rodents to humans, reinforcer devaluation has been
employed to study genuinely goal-directed choice in
humans, that is, to search for the underpinnings of beha-
vior that is flexibly adjusted to such things as changes in
the value of a predicted outcome. The orbitofrontal
cortex (OFC) has duly been revealed as playing a particu-
lar role in representing goal-directed value [26]. How-
ever, the bulk of experiments has not set out to
distinguish model-based from model-free systems, and
has rather more readily found regions implicated in
model-free processes. There is some indirect evidence
[27–29,30] for the involvement of dopamine and dopa-
minergic mechanisms in learning from reward prediction
errors in humans, along with more direct evidence from an
experiment involving the pharmacological manipulation
of dopamine [31]. These studies, in which prediction
errors have been explicitly modeled, along with others,
which use a more general multivariate approach [32] or an
argument based on a theoretical analysis of multiple,
separate, cortico-basal ganglia loops [33], also reveal roles
for OFC, medial prefrontal cortical structures, and even
the cingulate cortex. The contributions of the latter
especially have recently been under scrutiny in animal
studies focused on the cost-benefit tradeoffs inherent in
decision-making [34,35], and the involvement of dopa-
minergic projections to and from the anterior cingulate
cortex, and thus potential interactions with RL have been
suggested ([36], but see [37]). However, novel approaches
to distinguishing model-based and model-free control
may be necessary to tease apart more precisely the
singular contributions of the areas.
Computation of prediction errors: in terms of pathways, one
of the more striking recent findings is evidence that the
lateral habenula suppresses the activity of dopamine
neurons [38,39], in a way which may be crucial for the
representation of the negative prediction errors that arise
when states turn out to be worse than expected. One
natural possibility is then that pauses in the burst firing of
dopamine cells might code for these negative prediction
errors. This has received some quantitative support [40],
despite the low baseline rate of firing of these neurons,
which suggests that such a signal would have a rather low
bandwidth. Further, studies examining the relationship
between learning from negative and positive con-
sequences and genetic polymorphisms related to the
D2 dopaminergic receptor have established a functional
link between dopamine and learning when outcomes are
worse than expected [41,42]. Unfortunately, marring this
apparent convergence of evidence is the fact that the
distinction between the absence of an expected reward
and more directly aversive events is still far from being
clear, as we complain below.
Current Opinion in Neurobiology 2008, 18:185–196
Further data on contributions to the computation of the
TD prediction error have come from new findings on
excitatory pathways into the dopamine system [43]. Evi-
dence about the way that the amygdala [44,45,46] and
the medial prefrontal cortex [47] code for both positive
and negative predicted values and errors, and the joint
coding of actions, the values of actions, and rewards in
striatal and OFC activity [48–55,9], is also significant,
given the putative roles of these areas in learning and
representing various forms of RL values.
In fact, the machine learning literature has proposed
various subtly different versions of the TD learning
signal, associated with slightly different model-free RL
methods. Recent evidence from a primate study [56]
looking primarily at one dopaminergic nucleus, the sub-
stantia nigra pars compacta, seems to support a version
called SARSA [57]; whereas evidence from a rodent study
[58] of the other major dopaminergic nucleus, the ventral
tegmental area, favors a different version called Q-learn-
ing (CJCH Watkins, Learning from delayed rewards,
PhD thesis, University of Cambridge, 1989). Resolving
this discrepancy, and indeed, determining whether these
learning rules can be incorporated within the popular
Actor/Critic framework for model-free RL in the basal
ganglia [59,1], will necessitate further experiments and
computational investigation.
A more radical change to the rule governing the activity of
dopamine cells which separates out differently the por-
tions associated with the outcome (the primary reward)
and the learned predictions has also been suggested in a
modeling study [60]. However, various attractive features
of the TD rule, such as its natural account of secondary
conditioning and its resulting suitability for optimizing
sequences of actions leading up to a reward, are not
inherited directly by this rule.
Temporal discounting: a recurrent controversy involves the
way that the utilities of proximal and distant outcomes are
weighed against each other. Exponential discounting,
similar to a uniform interest rate, has attractive theoretical
properties, notably the absence of intertemporal choice
conflict, the possibility of recursive calculation scheme
and simple prediction errors [1]. However, the more
computationally complex hyperbolic discounting, which
shows preference reversals and impulsivity, is a more
common psychological finding in humans and animals
[61].
The immediate debate concerned the abstraction and
simplification of hyperbolic discounting to two evaluative
systems, one interested mostly in the here and now, the
other in the distant future [62], and their apparent instan-
tiation in different subcortical and cortical structures
respectively [63,64]. This idea became somewhat
wrapped-up with the distinct notion that these neural
www.sciencedirect.com
 Reinforcement learning: The Good, The Bad and The Ugly Dayan and Niv 189
areas are involved in model-free and model-based learn-
ing. Further, other studies found a more unitary neural
representation of discounting [65], at least in the BOLD
signal, and recent results confirm that dopaminergic pre-
diction errors indeed show the expected effects of dis-
counting [58]. One surprising finding is that the OFC
may separate out the representation of the temporal
discount factor applied to distant rewards from that of
the magnitude of the reward [54], implying a complex
problem of how these quantities are then integrated.
There has also been new work based on the theory [8]
that the effective interest rate for time is under the
influence of the neuromodulator serotonin. In a task that
provides a fine-scale view of temporal choice [66], dietary
reduction of serotonin levels in humans (tryptophan
depletion) gave rise to extra impulsivity, that is, favoring
smaller rewards sooner over larger rewards later, which
can be effectively modeled by a steeper interest rate [67].
Somewhat more problematically for the general picture of
control sketched above, tryptophan depletion also led to a
topographically mapped effect in the striatum, with
quantities associated with predictions for high interest
rates preferentially correlated with more ventral areas,
and those for low interest rates with more dorsal areas
[68].
The idea that subjects are trying to optimize their long-
run rates of acquisition of reward has become important in
studies of time-sensitive sensory decision-making [11,69].
It also inspired a new set of modeling investigations into
free operant choice tasks, in which subjects are free to
execute actions at times of their own choosing, and the
dependent variables are quantities such as the rates of
responding [70,71]. In these accounts, the rate of reward
acts as an opportunity cost for time, thereby penalizing
sloth, and is suggested as being coded by the tonic (as
distinct from the phasic) levels of dopamine. This cap-
tures findings associated with response vigor given dopa-
minergic manipulations [72,73,74], and, at least
assuming a particular coupling between phasic and tonic
dopamine, can explain results linking vigor to predictions
of reward [75,76].
‘The Bad’: apparent but tractable
inconsistencies
Various research areas which come in close contact with
different aspects of RL, help extend or illuminate it in not
altogether expected ways. These include issues of aver-
sive-appetitive interactions, exploration and novelty, a
range of phenomena important in neuroeconomics [77,78]
such as risk, Pavlovian-instrumental interactions, and also
certain new structural or architectural findings. The exist-
ence of multiple control mechanisms makes it challen-
ging to interpret some of these results unambiguously,
since rather little is known for sure about the interaction
between model-free and model-based systems, or the way
www.sciencedirect.com
the neural areas involved communicate, cooperate and
compete.
Appetitive–aversive interactions: one key issue that dogs
neural RL [79] is the coding of aversive rather than
appetitive prediction errors. Although dopamine neurons
are seemingly mostly inhibited by unpredicted punish-
ments [80,81], fMRI studies into the ventral striatum in
humans have produced mixed results, with aversive pre-
diction errors sometimes leading to above-baseline
BOLD [82,83], but other times below-baseline BOLD,
perhaps with complex temporal dynamics [84,23]. Dopa-
mine antagonists suppress the aversive prediction error
signal [85], and withdrawing (OFF) or administering
(ON) dopamine-boosting medication to patients suffer-
ing from Parkinson’s disease leads to boosted and sup-
pressed learning from negative outcomes, respectively
[86].
One study that set out to compare directly appetitive and
aversive prediction errors found a modest spatial separ-
ation in the ventral striatal BOLD signal [87], consistent
with various other findings about the topography of this
structure [88,89]; indeed, there is a similar finding in the
OFC [90]. However, given that nearby neurons in the
amygdala code for either appetitive or aversive outcomes
[44,45,46], fMRI’s spotlight may be too coarse to address
all such questions adequately.
Aversive predictions are perhaps more complex than
appetitive ones, owing to their multifaceted range of
effects, with different forms of contextual information
(such as defensive distance; [91]) influencing the choice
between withdrawal and freezing versus approach and
fighting. Like appetitive choice behavior, some of these
behavioral effects require vigorous actions, which we
discussed above in terms of tonic levels of dopamine
[70,71]. Moreover, aversive predictions can lead to active
avoidance, which then brings about appetitive predictions
associated with the achievement of safety. This latter
mechanism is implicated in conditioned avoidance
responses [92], and has also been shown to cause increases
in BOLD in the same part of the OFC that is also
activated by the receipt of rewards [93].
Novelty, uncertainty and exploration: the bulk of work in RL
focuses on exploitation, that is on using past experience to
optimize outcomes on the next trial or next period. More
ambitious agents seek also to optimize exploration, taking
account in their choices not only the benefits of known
(expected) future rewards, but also the potential benefits
of learning about unknown rewards and punishments (i.e.
the long-term gain to be harvested due to acquiring
knowledge about the values of different states). The
balancing act between these requires careful accounting
for uncertainty, in which the neuromodulators acetyl-
choline and norepinephrine have been implicated
Current Opinion in Neurobiology 2008, 18:185–196
190 Cognitive neuroscience
[94,95]. Uncertainty is also related to novelty, which
seems to involve dopaminergic areas and some of their
targets [96,97,98]. Uncertainty and novelty can support
different types of exploration, respectively, directed
exploration, in which exploratory actions are taken in
proportion to known uncertainty, and undirected explora-
tion, in which novel unknown parts of the environment
are uniformly explored. Tracking uncertainty has been
shown to involve the PFC in humans [99] and complex
patterns of activity in lateral and medial prefrontal cortex
of monkeys [100], and has also been associated with
functions of the amygdala [101].
Model-free and model-based systems in RL adopt rather
different approaches to exploration, although because
uncertainty generally favors model-based control [14],
the model-free approaches may be concealed. Indeed,
we may expect that certain forms of explicit change and
uncertainty could act to transfer habitized control back to
a model-based system [102]. Model-based exploration
probably involves frontal areas, as exemplified by one
recent imaging study which revealed a particular role for
fronto-polar cortex in trials in which nonexploitative
actions were chosen [103]. Model-free exploration has
been suggested to involve neuromodulatory systems such
as dopamine and norepinephrine, instantiating computa-
tionally more primitive strategies such as dopaminergic
exploration bonuses [98].
Uncertainty should influence learning as well as explora-
tion. In particular, learning (and forgetting) rates should
be higher in a rapidly changing environment and slower in
a rather stationary one. A recent fMRI study demon-
strated that human subjects adjust their learning rates
according to the volatility of the environment, and
suggested that the anterior cingulate cortex is involved
in the online tracking of volatility [104]. In animal con-
ditioning tasks involving a more discrete form of surprise
associated with stimuli, substantial studies have shown a
role for the central nucleus of the amygdala and the
cholinergic neurons in the substantia innominata and
nucleus basalis in upregulating subsequent learning
associated with that stimulus [94]. More recent work
has shown that there is a crucial role for the projection
from the dopaminergic (putatively prediction error cod-
ing) substantia nigra pars compacta to the central nucleus
in this upregulation [105], and that the involvement of the
central nucleus is limited to the time of surprise (and that
of the cholinergic nuclei to the time of subsequent
learning; [106]). This accords with evidence that some
amygdala neurons respond directly to both positive and
negative surprising events ([45], a minimum require-
ment for a surprise signal) and to unpredictability itself
[101].
Risk, regret and neuroeconomics: the relatively straightfor-
ward view of utility optimization that permeates neural
Current Opinion in Neurobiology 2008, 18:185–196
RL is challenged by various findings in and around the
fields of experimental economics, behavioral economics,
and neuroeconomics [107,77]. These study the psycho-
logical and neural factors underlying a range of situations
in which behavior departs from apparent normative
ideals. One example of this concerns the sensitivity of
choice behavior to risk associated with variability in the
outcomes. There are extensive economic theories about
risk, and indeed evidence from functional neuroimaging
that variability may be coded in specific cortical and
subcortical regions such as the insular cortex, OFC and
the ventral striatum [108–112]. However, how risk should
influence action selection in model-free and model-based
control is not completely clear, partly because variability
can have both an indirect influence on values, for instance
via a saturating utility function for rewards [113], and
direct effects through a risk-sensitive learning process
[114,115]. Although nonlinear utilities for rewards will
affect both types of controllers, the effects of risk through
sampling biases [114] or through a learning rule that is
asymmetrically sensitive to positive and negative errors
[115] depend on the specific implementation of learning
in a model-free or model-based controller, and thus can
differ between the two.
A second neuroeconomic concern is regret, in which a form
of counterfactual thinking [116] or fictive learning [117] is
induced when foregone alternatives turn out to be better
than those that were chosen [118,119]. Imaging studies
have suggested a role for the OFC and other targets of the
dopamine system in processing fictive learning signals
[120,121,117]. Here, also, there may be separate instan-
tiations in model-based and model-free systems, as
accounting for counterfactuals is straightforward in the
former, but requires extra machinery in the latter.
A third issue along these lines is framing, in which
different descriptions of a single (typically risky) outcome
result in different valuations. Subjects are more reluctant
to take chances in the face of apparent gains than losses,
even when the options are actually exactly the same. A
recent imaging study [122] reported a close correlation
between susceptibility to framing and activity in the
amygdala, and a negative correlation with activity in
OFC and medial PFC.
Finally, there is an accumulating wealth of work on social
utility functions, and game–theoretic interactions be-
tween subjects [123], including studies in patient popu-
lations with problems with social interactions [124].
There are interesting hints for neural RL from studies
that use techniques such as transcranial magnetic stimu-
lation to disrupt specific prefrontal activity, which turns
out to affect particular forms of social choice such as the
ability to reject patently unfair (though still lucrative)
offers in games of economic exchange [125]. However,
these have yet to be coupled to the nascent more com-
www.sciencedirect.com
 Reinforcement learning: The Good, The Bad and The Ugly Dayan and Niv 191
plete RL accounts of the complex behavior in pair and
multi-agent tournaments [126,127].
Pavlovian values: although it is conventional in RL to
consider Pavlovian or classical conditioning as only being
about the acquisition of predictions, with instrumental
conditioning being wholly responsible for choice, it is
only through a set of (evolutionarily programmed)
responses, such as the approach that is engendered by
predictions of reward, that Pavlovian predictions become
evident in the first place. Indeed, Pavlovian responses can
famously out-compete instrumental responses [128]. This
is of particular importance in omission schedules or nega-
tive automaintenance, when subjects continue respond-
ing based on predictions of future rewards, even when
their actions actually prevent them from getting those
rewards.
Other paradigms such as Pavlovian-instrumental transfer,
in which Pavlovian predictors influence the vigor of
instrumental responding, are further evidence of non-
normative interactions among these two forms of learn-
ing, potentially accounting for some of the effects of
manipulating reward schedules on effort [52,51]. Part
of the architecture of this transfer involving the amgydala,
the striatum and dopamine that has been worked out in
rodent studies [17] has recently been confirmed in a
human imaging experiment [129]; however, a recent
surprise was that lesions in rats that disconnected the
central nucleus of the amygdala from the ventral teg-
mental area actually enhanced transfer [130], rather than
suppressing it, as would have been expected from earlier
work (e.g. [131]).
The Pavlovian preparatory response to aversive predic-
tions is often withdrawal or disengagement. It has been
hypothesized that this is a form of serotonergically
mediated inhibition [132,133] that acts to arrest paths
leading to negative outcomes, and, in model-based terms,
excise potentially aversive parts of the search tree [134].
Compromising serotonin should thus damage this reflex-
ive inhibition, leading to systematic problems with evalu-
ation and choice. The consummatory Pavlovian response
to immediate threat is panic, probably mediated by the
peri-acqueductal gray [135,91]; such responses have
themselves recently been observed in an imaging study
which involved a virtual threatening ‘chase’ [136].
Appetitive and aversive Pavlovian influences have been
suggested as being quite pervasive [137], accounting for
aspects of impulsivity apparent in hyperbolic discounting,
framing and the like. Unfortunately, the sophisticated
behavioral and neural analyses of model-free and model-
based instrumental values are not paralleled, as of yet, by
an equivalently worked-out theory for the construction of
Pavlovian values. Moreover, Pavlovian influences may
affect model-based and model-free instrumental actions
www.sciencedirect.com
differently, suggesting even more relatively untapped
complexity.
Structural findings: finally, there are some recent fascinat-
ing and sometimes contrarian studies into the way that
striatal pathways function in choice. One set
[138,139,140] has revealed a role for the subthalamic
nucleus (STN) in slowing down choices between strongly
appetitive options, perhaps to give more time for the
precise discrimination of the best action from merely
good ones [141]. It could again be that the impulsivity
created when the STN is suppressed arises from a form of
Pavlovian approach [137]. Others have looked at serial
processes that shift action competition and choice up a
ventral to dorsal axis along the striatum [142,143], perhaps
consistent with the suggested spiraling connectivity
through dopaminergic nuclei [144,145].
‘The Ugly’: crucial challenges
The last set of areas of neural RL suffer a crucial disparity
between their importance and the relative dearth of
systematic or comprehensive studies. Hopefully, by the
next review, this imbalance will be at least partially
redressed.
In terms of model-based control, a central question con-
cerns the acquisition and use of hierarchical structures. It
seems rather hopeless to plan using only the smallest
units of action (think of the twitches of a single muscle),
and so there is a great interest in hierarchies in both
standard [146] and neural RL [147]. Unfortunately, the
crucial problem of acquiring appropriate hierarchical
structure in the absence of supervision is far from being
solved. A second concern has to do with a wide class of
learning scenarios in which optimal learning relies on
detection of change and, essentially, learning of a ‘new’
situation, rather than updating the previously learned
one. A prime example of this is the paradigm of extinction
in which a cue previously predictive of reward is no longer
paired with the rewarding outcome. Behavioral results
show that animals and humans do not simply ‘unlearn’ the
previous predictive information, but rather they learn a
new predictive relationship that is inhibitory to the old
one [148]. How this type of learning is realized within a
RL framework, whether model-based or model-free, is at
present unclear, though some suggestions have been
made [149]. The issues mentioned above to do with
uncertainty and change also bear on this.
In terms of model-free control, there are various pressing
anomalies (on top of appetitive/aversive opponency)
associated with the activity of dopamine cells (and the
not-necessarily equivalent release of dopamine at its
targets; [150]). One concerns prediction errors inspired
by conditioned inhibitors for reward and punishment
predictors. While the predictive stimuli we have contem-
plated so far are stimuli that predict the occurrence of
Current Opinion in Neurobiology 2008, 18:185–196
192 Cognitive neuroscience
some outcome, a conditioned inhibitor is a stimulus that
predicts that an otherwise expected outcome will not
happen. How the former (conditioned excitors) and the
latter (conditioned inhibitors) interact to generate a single
prediction and a corresponding prediction error is not yet
clear either neurally or computationally. A fascinating
electrophysiological study [151] targeting this issue
suggests that the answer is not simple.
A second anomaly is the apparent adaptive scaling of
prediction errors depending on the precise range of such
errors that is expected, forming an informationally effi-
cient code [152]. As with other cases of adaptation, it is not
clear that downstream mechanisms (here, presumably
dopamine receptors) have the information or indeed
the calculational wherewithal to reverse engineer cor-
rectly the state of adaptation of the processes controlling
the activity of dopamine cells. This opens up the possib-
ility that there might be biases in the interpretation of the
signal, leading to biases in choice. Alternatively, this
adaptation might serve a different computational goal,
such as adjusting learning appropriately in the face of
forms of uncertainty [153].
A timely reminder of the complexity of the dopamine
system came from a recent study suggesting a new sub-
class of dopamine neurons with particularly unusual neu-
rophysiological properties, that selectively target the
prefrontal cortex, the basolateral amygdala and the core
of the accumbens [154]. It is important to remember that
most of our knowledge about the behaviorally relevant
activity of dopamine neurons comes from recordings
during Pavlovian tasks or rather simple instrumental
scenarios — basic questions about dopaminergic firing
patterns when several actions are necessary in order to
obtain an outcome, when there are several outcomes in a
trial, or in hierarchical tasks are still unanswered. Lammel
et al.’s [154] results add to previous subtle challenges to
the common premise that the dopaminergic signal is
unitary (e.g. when taking together [56,58]). Future
physiological studies and recordings in complex tasks
are needed to flesh out what could potentially be a more
diverse signal than has so far been considered. The role of
dopamine in the prefrontal cortex, which has only played
prominently in relatively few models (notably [155]), and
indeed its potential effects on the model-based controller,
might also need to be rethought in light of these findings.
A final area of unfortunate lack of theory and dearth of
data has to do with timing. The activity pattern of
dopamine cells that most convincingly suggests that they
report a prediction error is the well-timed dip in respond-
ing when an expected reward is not provided [156,40].
The neural mechanisms underlying this timing are most
unclear; in fact there are various quite different classes of
suggestion in the RL literature and beyond. What is worse
is that this form of interval timing is subject to substantial
Current Opinion in Neurobiology 2008, 18:185–196
scalar noise [157], to a degree that could make accurate
prediction learning, and accurately timed dips in dopa-
mine responding, quite challenging to achieve.
Conclusions
As should be apparent from this review, neural RL is a
vibrant and dynamic field, generating new results at a near-
overwhelming rate, and spreading its wings well beyond its
initial narrow confines of trial-and-error reward learning.
We have highlighted many foci of ongoing study, and also
some orphaned areas mentioned in the previous section.
However, our best hope is that the sterling efforts to link
together the substantial theoretically motivated and infor-
mative animal studies to human neuroimaging results,
along with new data from cyclic voltammetric measure-
ments of phasic dopamine concentrations [158], imminent
results on serotonin, and even nascent efforts to activate
DA cells in vivo using new optogenetic methods such as
targetted channel rhodopsin, will start to pay off in the
opposite direction by deciding between, and forcing
improvements to, RL models.
Acknowledgements
We are very grateful to Peter Bossaerts, Nathaniel Daw, Michael Frank,
Russell Poldrack, Daniel Salzman, Ben Seymour and Wako Yoshida for their
helpful comments on a previous version of this manuscript. PD is funded by
the Gatsby Charitable Foundation; YN is funded by the Human Frontiers
Science Program.
References and recommended reading
Papers of particular interest, published within the period of the review,
have been highlighted as:
 of special interest
 of outstanding interest
1. Sutton RS, Barto AG: Reinforcement Learning: An Introduction
(Adaptive Computation and Machine Learning)The MIT Press;
1998.
2. Montague PR: Why Choose This Book?: How We Make
Decisions.Dutton Adult; 2006.
3. Sutton R: Learning to predict by the methods of temporal
differences. Mach Learn 1988, 3:9-44.
4. Montague PR, Dayan P, Sejnowski TJ: A framework for
mesencephalic dopamine systems based on predictive
Hebbian learning. J Neurosci 1996, 16:1936-1947.
5. Daw ND, Doya K: The computational neurobiology of learning
and reward. Curr Opin Neurobiol 2006, 16:199-204.
6. Johnson A, van der Meer MA, Redish AD: Integrating
hippocampus and striatum in decision-making. Curr Opin
Neurobiol 2007, 17:692-697.
7. O’Doherty JP, Hampton A, Kim H: Model-based fMRI and its
application to reward learning and decision making. Ann N Y
Acad Sci 2007, 1104:35-53.
8. Doya K: Modulators of decision making. Nat Neurosci 2008,
11:410-416.
9. Rushworth MFS, Behrens TEJ: Choice, uncertainty and value in
prefrontal and cingulate cortex. Nat Neurosci 2008, 11:389-397.
10. Körding K: Decision theory: what ‘‘should’’ the nervous system
do? Science 2007, 318:606-610.
11. Gold JI, Shadlen MN: The neural basis of decision making. Annu
Rev Neurosci 2007, 30:535-574.
www.sciencedirect.com
 Reinforcement learning: The Good, The Bad and The Ugly Dayan and Niv 193
12. Lee D: Neural basis of quasi-rational decision making. Curr
Opin Neurobiol 2006, 16:191-198.
13. Niv Y, Montague PR: Theoretical and empirical studies of
learning. In Neuroeconomics: Decision Making and The Brain.
Edited by Glimcher PW, Camerer C, Fehr E, Poldrack R. New York,
NY: Academic Press; 2008:329–349.
14. Daw ND, Niv Y, Dayan P: Uncertainty-based competition
between prefrontal and dorsolateral striatal systems for
behavioral control. Nat Neurosci 2005, 8:1704-1711.
15. Dickinson A, Balleine B: The role of learning in motivation. In
Stevens’ Handbook of Experimental Psychology. Edited by
Gallistel C. New York, NY: Wiley; 2002:497–533.
16. Uylings HBM, Groenewegen HJ, Kolb B: Do rats have a
prefrontal cortex? Behav Brain Res 2003, 146:3-17.
17. Balleine BW: Neural bases of food-seeking: affect, arousal and
reward in corticostriatolimbic circuits. Physiol Behav 2005,
86:717-730.
18. Killcross S, Coutureau E: Coordination of actions and habits in
the medial prefrontal cortex of rats. Cereb Cortex 2003, 13:400-
408.
19. Dolan RJ: The human amygdala and orbital prefrontal cortex in
behavioural regulation. Philos Trans R Soc Lond B: Biol Sci 2007,
362:787-799.
20. Matsumoto K, Tanaka K: The role of the medial prefrontal cortex
in achieving goals. Curr Opin Neurobiol 2004, 14:178-185.
21. Baxter J, Bartlett P: Infinite-horizon policy-gradient estimation.
J Artif Intell Res 2001, 15:319-350.
22. Berns GS, McClure SM, Pagnoni G, Montague PR: Predictability
modulates human brain response to reward. J Neurosci 2001,
21:2793-2798.
23. O’Doherty JP, Dayan P, Friston K, Critchley H, Dolan RJ:
Temporal difference models and reward-related learning in
the human brain. Neuron 2003, 38:329-337.
24. Haruno M, Kuroda T, Doya K, Toyama K, Kimura M, Samejima K,
Imamizu H, Kawato M: A neural correlate of reward-based
behavioral learning in caudate nucleus: a functional magnetic
resonance imaging study of a stochastic decision task. J
Neurosci 2004, 24:1660-1665.
25. Logothetis NK, Wandell BA: Interpreting the BOLD signal. Annu
Rev Physiol 2004, 66:735-769.
26. Valentin VV, Dickinson A, O’Doherty JP: Determining the neural
 substrates of goal-directed learning in the human brain. J
Neurosci 2007, 27:4019-4026. Rodent behavioral paradigms [15]
have played a crucial role in establishing the structural and
functional dissociations between different classes of behavior [14].
This paper is one of a series of studies from this group which is
importing these paradigms lock, stock and barrel to humans,
showing that exactly the same distinctions and indeed
homologous anatomical bases apply. These studies form a crucial
part of the evidentiary basis of neural RL.
27. O’Doherty JP, Buchanan TW, Seymour B, Dolan RJ: Predictive
neural coding of reward preference involves dissociable
responses in human ventral midbrain and ventral striatum.
Neuron 2006, 49:157-166.
28. Schönberg T, Daw ND, Joel D, O’Doherty JP: Reinforcement
learning signals in the human striatum distinguish learners
from nonlearners during reward-based decision making. J
Neurosci 2007, 27:12860-12867.
29. Tobler PN, O’Doherty JP, Dolan RJ, Schultz W: Human neural
learning depends on reward prediction errors in the blocking
paradigm. J Neurophysiol 2006, 95:301-310.
30. D’Ardenne K, McClure SM, Nystrom LE, Cohen JD: Bold
 responses reflecting dopaminergic signals in the human
ventral tegmental area. Science 2008, 319:1264-1267. This
study achieves an impressive new level of anatomical precision in
recording fMRI BOLD signals from the ventral tegmental area (VTA)
in humans, confirming various aspects suggested from recordings
in macaques and rats (though the extra precision does not extend
to a distinction between dopaminergic and nondopaminergic cells
www.sciencedirect.com
in the VTA, let alone the different classes of dopamine cells). It will
be fascinating to see what happens in the substantia nigra pars
compacta, whose dopamine cells may be more closely associated
with action rather than value learning..
31. Pessiglione M, Seymour B, Flandin G, Dolan RJ, Frith CD:
 Dopamine-dependent prediction errors underpin
reward-seeking behaviour in humans. Nature 2006,
442:1042-1045. Despite the substantial circumstantial data from
animals and human neurological cases, there are few direct tests
of the specifically dopaminergic basis of normal instrumental
conditioning. This study tested this in healthy subjects using L-
DOPA and haloperidol, which are expected respectively to
enhance and suppress dopamine function. L-DOPA improved
choice performance toward monetary gains (compared to
haloperidol-treated subjects) but not avoidance of monetary
losses. This was accompanied by an increase in ventral striatal
BOLD responses corresponding to prediction errors in the gain
condition in the L-DOPA group, whose magnitude was sufficient
for a standard action-value learning model to explain the effects of
the drugs on behavioral choices. These results provide (much
needed, but still indirect) evidence that the ventral striatal
prediction error BOLD signal indeed reflects (at least partly)
dopaminergic activity.
32. Hampton AN, O’Doherty JP: Decoding the neural substrates of
reward-related decision making with functional MRI. Proc Natl
Acad Sci U S A 2007, 104:1377-1382.
33. Samejima K, Doya K: Multiple representations of belief states
and action values in corticobasal ganglia loops. Ann NY Acad
Sci 2007, 1104:213-228.
34. Walton ME, Bannerman DM, Alterescu K, Rushworth MFS:
Functional specialization within medial frontal cortex of the
anterior cingulate for evaluating effort-related decisions. J
Neurosci 2003, 23:6475-6479.
35. Schweimer J, Hauber W: Involvement of the rat anterior
cingulate cortex in control of instrumental responses
guided by reward expectancy. Learn Mem 2005,
12:334-342.
36. Schweimer J, Hauber W: Dopamine D1 receptors in the anterior
cingulate cortex regulate effort-based decision making. Learn
Mem 2006, 13:777-782.
37. Walton ME, Croxson PL, Rushworth MFS, Bannerman DM: The
mesocortical dopamine projection to anterior cingulate cortex
plays no role in guiding effort-related decisions. Behavioral
Neuroscience 2005, 119:323-328.
38. Matsumoto M, Hikosaka O: Lateral habenula as a source of
 negative reward signals in dopamine neurons. Nature 2007,
447:1111-1115. The sources of excitation and inhibition of
dopaminergic and indeed nondopaminergic cells in the VTA have
long been debated. This paper shows that there is an important
pathway from a nucleus called the lateral habenula, that has the
effect of inhibiting the activity of the dopamine cells. Consistent
with this connection, habenula neurons were excited by stimuli
that effectively predict the absence of reward, and inhibited by
targets that predict its presence.
39. Lecourtier L, Defrancesco A, Moghaddam B: Differential tonic
influence of lateral habenula on prefrontal cortex and nucleus
accumbens dopamine release. European Journal Neuroscience
2008, 27:1755-1762.
40. Bayer HM, Lau B, Glimcher PW: Statistics of midbrain dopamine
neuron spike trains in the awake primate. J Neurophysiol 2007,
98:1428-1439.
41. Frank MJ, Moustafa AA, Haughey HM, Curran T, Hutchison KE:
Genetic triple dissociation reveals multiple roles for dopamine
in reinforcement learning. Proc Natl Acad Sci U S A 2007,
104:16311-16316.
42. Klein TA, Neumann J, Reuter M, Hennig J, von Cramon DY,
Ullsperger M: Genetically determined differences in learning
from errors. Science 2007, 318:1642-1645.
43. McHaffie JG, Jiang H, May PJ, Coizet V, Overton PG, Stein BE,
Redgrave P: A direct projection from superior colliculus to
substantia nigra pars compacta in the cat. Neuroscience 2006,
138:221-234.
Current Opinion in Neurobiology 2008, 18:185–196
194 Cognitive neuroscience
44. Paton JJ, Belova MA, Morrison SE, Salzman CD: The primate
amygdala represents the positive and negative value of visual
stimuli during learning. Nature 2006, 439:865-870.
45. Belova MA, Paton JJ, Morrison SE, Salzman CD: Expectation
 modulates neural responses to pleasant and aversive stimuli
in primate amygdale. Neuron 2007, 55:970-984 The amygdala
has long been suggested to be an important site for appetitive and
aversive learning (in which the difference between valences is
crucial) and also a form of uncertainty (in which positive and
negative valences are treated together). This study and its
colleague [44] provides an interesting view into the coding of
valenced and unvalenced predictions and prediction errors in the
macaque amygdala. Putting these data together with the rodent
studies, with their anatomically and functionally precise
dissociations [17] is a pressing task..
46. Salzman CD, Paton JJ, Belova MA, Morrison SE: Flexible neural
representations of value in the primate brain. Ann N Y Acad Sci
2007, 1121:336-354.
47. Matsumoto M, Matsumoto K, Abe H, Tanaka K: Medial prefrontal
cell activity signaling prediction errors of action values. Nat
Neurosci 2007, 10:647-656.
48. Balleine BW, Delgado MR, Hikosaka O: The role of the dorsal
striatum in reward and decision-making. J Neurosci 2007,
27:8161-8165.
49. Hikosaka O: Basal ganglia mechanisms of reward-oriented eye
movement. Ann N Y Acad Sci 2007, 1104:229-249.
50. Lau B, Glimcher PW: Action and outcome encoding in the
primate caudate nucleus. J Neurosci 2007, 27:14502-14514.
51. Hikosaka O, Nakamura K, Nakahara H: Basal ganglia orient eyes
to reward. J Neurophysiol 2006, 95:567-584.
52. Simmons JM, Ravel S, Shidara M, Richmond BJ: A comparison of
reward-contingent neuronal activity in monkey orbitofrontal
cortex and ventral striatum: guiding actions toward rewards.
Ann NY Acad Sci 2007, 1121:376-394.
53. Padoa-Schioppa C: Orbitofrontal cortex and the computation
of economic value. Ann N Y Acad Sci 2007, 1121:232-253.
54. Roesch MR, Taylor AR, Schoenbaum G: Encoding of time-
discounted rewards in orbitofrontal cortex is independent of
value representation. Neuron 2006, 51:509-520.
55. Furuyashiki T, Holland PC, Gallagher M: Rat orbitofrontal cortex
separately encodes response and outcome information
during performance of goal-directed behaviour. J Neurosci
2008, 28:5127-5138.
56. Morris G, Nevet A, Arkadir D, Vaadia E, Bergman H: Midbrain
 dopamine neurons encode decisions for future action. Nat
Neurosci 2006, 9:1057-1063 Most historical recordings from
dopamine cells have been in cases in which there is no meaningful
choice between different alternatives. This study in macaques,
together with Roesch et al’s [58] in rodents looks specifically at
choice. The two studies confirm the general precepts of neural RL
based on temporal difference (TD) learning, but use rather different
methods, and come to different conclusions about which of the
variants of TD is at work..
57. Rummery G, Niranjan M: On-line Q-learning using connectionist
systems, Tech. Rep. Technical Report CUED/F-INFENG/TR 166,
Cambridge University Engineering Department, 1994.
58. Roesch MR, Calu DJ, Schoenbaum G: Dopamine neurons encode
 the better option in rats deciding between differently delayed or
sized rewards. Nat Neurosci 2007, 10:1615-1624 Along with work
from Hyland and colleagues [159,160], this is one of the few studies
looking at the activity of dopamine neurons in rodents. Many of its
findings are exactly consistent with neural TD, though it comes to a
different conclusion about exactly which TD rule is operational from
the monkey experiment performed by Morris et al. [56]..
59. Barto A: Adaptive critics and the basal ganglia. In Models of
Information Processing in the Basal Ganglia. Edited by Houk J,
Davis J, Beiser D. Cambridge, MA: MIT Press; 1995:215–232.
60. O’Reilly RC, Frank MJ, Hazy TE, Watz B: PVLV: the primary value
and learned value Pavlovian learning algorithm. Behav
Neurosci 2007, 121:31-49.
Current Opinion in Neurobiology 2008, 18:185–196
61. Ainslie G: Breakdown of Will. Cambridge University Press; 2001.
62. Loewenstein G, Prelec D: Anomalies in intertemporal choice:
Evidence and an interpretation. The Quarterly Journal of
Economics 1992, 107:573-597.
63. McClure SM, Laibson DI, Loewenstein G, Cohen JD: Separate
neural systems value immediate and delayed monetary
rewards. Science 2004, 306:503-507.
64. McClure SM, Ericson KM, Laibson DI, Loewenstein G, Cohen JD:
Time discounting for primary rewards. J Neurosci 2007,
27:5796-5804.
65. Kable JW, Glimcher PW: The neural correlates of subjective
value during intertemporal choice. Nat Neurosci 2007,
10:1625-1633.
66. Schweighofer N, Shishida K, Han CE, Okamoto Y, Tanaka SC,
Yamawaki S, Doya K: Humans can adopt optimal discounting
strategy under real-time constraints. PLoS Comput Biol 2006,
2:e152.
67. Schweighofer N, Bertin M, Shishida K, Okamoto Y, Tanaka SC,
Yamawaki S, Doya K: Low-serotonin levels increase delayed
reward discounting in humans. J Neurosci 2008, 28:4528-4532.
68. Tanaka SC, Schweighofer N, Asahi S, Shishida K, Okamoto Y,
Yamawaki S, Doya K: Serotonin differentially regulates short-
and long-term prediction of rewards in the ventral and dorsal
striatum. PLoS ONE 2007, 2:e1333.
69. Bogacz R, Brown E, Moehlis J, Holmes P, Cohen JD: The physics
of optimal decision making: a formal analysis of models of
performance in two-alternative forced-choice tasks. Psychol
Rev 2006, 113:700-765.
70. Niv Y, Joel D, Dayan P: A normative perspective on motivation.
Trends Cogn Sci 2006, 10:375-381.
71. Niv Y, Daw ND, Joel D, Dayan P: Tonic dopamine: opportunity
costs and the control of response vigor. Psychopharmacology
(Berl) 2007, 191:507-520.
72. Salamone JD, Correa M: Motivational views of reinforcement:
implications for understanding the behavioral functions of
nucleus accumbens dopamine. Behav Brain Res 2002, 137:3-25.
73. Nakamura K, Hikosaka O: Role of dopamine in the primate
caudate nucleus in reward modulation of saccades. J Neurosci
2006, 26:5360-5369.
74. Mazzoni P, Hristova A, Krakauer JW: Why don’t we move faster?
 Parkinson’s disease, movement vigor, and implicit motivation.
J Neurosci 2007, 27:7105-7116 Patients with Parkinson’s disease
exhibit bradykinesia, that is, they move more slowly than healthy
controls. This paper suggests that this effect is not from the
obvious cause of an altered speed-accuracy tradeoff, but rather
because the patient’s ‘motivation to move’, defined in a careful
way that resonates well with Niv et al’s [71] notion of vigor, is
reduced..
75. Pessiglione M, Schmidt L, Draganski B, Kalisch R, Lau H,
Dolan RJ, Frith CD: How the brain translates money into force: a
neuroimaging study of subliminal motivation. Science 2007,
316:904-906.
76. Shidara M, Richmond BJ: Differential encoding of information
about progress through multi-trial reward schedules by three
groups of ventral striatal neurons. Neurosci Res 2004, 49:307-
314.
77. Glimcher PW: Decisions, Uncertainty, and the Brain: The Science
of Neuroeconomics (Bradford Books). The MIT Press; 2004.
78. Montague PR: Neuroeconomics: a view from neuroscience.
Funct Neurol 2007, 22:219-234.
79. Daw ND, Kakade S, Dayan P: Opponent interactions between
serotonin and dopamine. Neural Netw 2002, 15:603-616.
80. Ungless MA, Magill PJ, Bolam JP: Uniform inhibition of
dopamine neurons in the ventral tegmental area by aversive
stimuli. Science 2004, 303:2040-2042.
81. Coizet V, Dommett EJ, Redgrave P, Overton PG: Nociceptive
responses of midbrain dopaminergic neurones are modulated
www.sciencedirect.com
 Reinforcement learning: The Good, The Bad and The Ugly Dayan and Niv 195
by the superior colliculus in the rat. Neuroscience 2006,
139:1479-1493.
82. Seymour B, O’Doherty JP, Dayan P, Koltzenburg M, Jones AK,
Dolan RJ, Friston KJ, Frackowiak RS: Temporal difference
models describe higher-order learning in humans. Nature
2004, 429:664-667.
83. Jensen J, Smith AJ, Willeit M, Crawley AP, Mikulis DJ, Vitcu I,
Kapur S: Separate brain regions code for salience vs. valence
during reward prediction in humans. Hum Brain Mapp 2007,
28:294-302.
84. Delgado MR, Nystrom LE, Fissell C, Noll DC, Fiez JA: Tracking the
hemodynamic responses to reward and punishment in the
striatum. J Neurophysiol 2000, 84:3072-3077.
85. Menon M, Jensen J, Vitcu I, Graff-Guerrero A, Crawley A,
Smith MA, Kapur S: Temporal difference modeling of the blood-
oxygen level dependent response during aversive
conditioning in humans: effects of dopaminergic modulation.
Biol Psychiatry 2007, 62:765-772.
86. Frank MJ, Seeberger LC, O’Reilly RC: By carrot or by stick:
cognitive reinforcement learning in Parkinsonism. Science
2004, 306:1940-1943.
87. Seymour B, Daw N, Dayan P, Singer T, Dolan R: Differential
encoding of losses and gains in the human striatum. J Neurosci
2007, 27:4826-4831.
88. Reynolds SM, Berridge KC: Fear and feeding in the nucleus
accumbens shell: rostrocaudal segregation of GABA-elicited
defensive behavior versus eating behaviour. J Neurosci 2001,
21:3261-3270.
89. Reynolds SM, Berridge KC: Positive and negative motivation in
nucleus accumbens shell: bivalent rostrocaudal gradients for
GABA-elicited eating, taste ‘‘liking’’/‘‘disliking’’ reactions,
place preference/avoidance, and fear. J Neurosci 2002,
22:7308-7320.
90. O’Doherty J, Kringelbach ML, Rolls ET, Hornak J, Andrews C:
Abstract reward and punishment representations in the
human orbitofrontal cortex. Nat Neurosci 2001, 4:95-102.
91. McNaughton N, Corr PJ: A two-dimensional neuropsychology
of defense: Fear/anxiety and defensive distance. Neurosci
Biobehav Rev 2004, 28:285-305.
92. Moutoussis M, Williams J, Dayan P, Bentall RP: Persecutory
delusions and the conditioned avoidance paradigm: towards
an integration of the psychology and biology of paranoia.
Cognit Neuropsychiatry 2007, 12:495-510.
93. Kim H, Shimojo S, O’Doherty JP: Is avoiding an aversive
outcome rewarding? Neural substrates of avoidance learning
in the human brain. PLoS Biol 2006, 4:e233.
94. Holland P, Gallagher M: Amygdala circuitry in attentional and
representational processes. Trends Cogn Sci 1999, 3:65-73.
95. Cohen JD, McClure SM, Yu AJ: Should I stay or should I go? How
the human brain manages the trade-off between exploitation
and exploration. Philos Trans R Soc Lond B Biol Sci 2007,
362:933-942.
96. Wittmann BC, Bunzeck N, Dolan RJ, Düzel E: Anticipation of
 novelty recruits reward system and hippocampus while
promoting recollection. Neuroimage 2007, 38:194-202 Novelty
plays a crucial role in learning, and has also been associated with
the dopamine system via the construct of bonuses [98]. This paper
is part of a series of studies into the coupling between novelty,
putatively dopaminergic midbrain activation and hippocampal-
dependent long-term memory. Along with their intrinsic interest,
the results of these experiments suggest that this form of learning
can provide evidence about the characteristics of signals
associated with RL..
97. Bunzeck N, Düzel E: Absolute coding of stimulus novelty in the
human substantia nigra/VTA. Neuron 2006, 51:369-379.
98. Kakade S, Dayan P: Dopamine: generalization and bonuses.
Neural Netw 2002, 15:549-559.
99. Yoshida W, Ishii S: Resolution of uncertainty in prefrontal
cortex. Neuron 2006, 50:781-789.
www.sciencedirect.com
100. Matsumoto M, Matsumoto K, Tanaka K: Effects of novelty on
activity of lateral and medial prefrontal neurons. Neurosci Res
2007, 57:268-276.
101. Herry C, Bach DR, Esposito F, Salle FD, Perrig WJ, Scheffler K,
Lthi A, Seifritz E: Processing of temporal unpredictability in
human and animal amygdale. J Neurosci 2007, 27:5958-5966.
102. Li J, McClure SM, King-Casas B, Montague PR: Policy adjustment
in a dynamic economic game. PLoS ONE 2006, 1:e103.
103. Daw ND, O’Doherty JP, Dayan P, Seymour B, Dolan RJ: Cortical
substrates for exploratory decisions in humans. Nature 2006,
441:876-879.
104. Behrens TEJ, Woolrich MW, Walton ME, Rushworth MFS:
Learning the value of information in an uncertain world. Nat
Neurosci 2007, 10:1214-1221.
105. Lee HJ, Youn JM, O MJ, Gallagher M, Holland PC. Role of
substantia nigra-amygdala connections in surprise-induced
enhancement of attention. J Neurosci 2006, 26:6077–6081.
106. Holland PC, Gallagher M: Different roles for amygdala central
nucleus and substantia innominata in the surprise-induced
enhancement of learning. J Neurosci 2006, 26:3791-3797.
107. Camerer C: Behavioral Game Theory: Experiments in Strategic
Interaction Princeton, NJ: Princeton University Press; 2003.
108. Kuhnen CM, Knutson B: The neural basis of financial risk taking.
Neuron 2005, 47:763-770.
109. Dreher JC, Kohn P, Berman KF: Neural coding of distinct
statistical properties of reward information in humans. Cereb
Cortex 2006, 16:561-573.
110. Tanaka SC, Samejima K, Okada G, Ueda K, Okamoto Y,
Yamawaki S, Doya K: Brain mechanism of reward prediction
under predictable and unpredictable environmental dynamics.
Neural Netw 2006, 19:1233-1241.
111. Tobler PN, O’Doherty JP, Dolan RJ, Schultz W: Reward value
coding distinct from risk attitude-related uncertainty coding in
human reward systems. J Neurophysiol 2007, 97:1621-1632.
112. Preuschoff K, Bossaerts P, Quartz SR: Neural differentiation of
expected reward and risk in human subcortical structures.
Neuron 2006, 51:381-390.
113. Tobler PN, Fletcher PC, Bullmore ET, Schultz W: Learning-related
human brain activations reflecting individual finances. Neuron
2007, 54:167-175.
114. Niv Y, Joel D, Meilijson I, Ruppin E: Evolution of reinforcement
learning in uncertain environments: a simple explanation for
complex foraging behaviors. Adaptive Behavior 2002, 10:5-24.
115. Mihatsch O, Neuneier R: Risk-sensitive reinforcement learning.
Mach Learn 2002, 49:267-290.
116. Byrne R: Mental models and counterfactual thoughts about
what might have been. Trends Cogn Sci 2002, 6:426-431.
117. Lohrenz T, McCabe K, Camerer CF, Montague PR: Neural
signature of fictive learning signals in a sequential investment
task. Proc Natl Acad Sci U S A 2007, 104:9493-9498.
118. Bell D: Regret in decision making under uncertainty. Oper Res
1982, 30:961-981.
119. Loomes G, Sugden R: Regret theory: an alternative theory of
rational choice under uncertainty. Econ J 1982, 92:805-824.
120. Breiter HC, Aharon I, Kahneman D, Dale A, Shizgal P: Functional
imaging of neural responses to expectancy and experience of
monetary gains and losses. Neuron 2001, 30:619-639.
121. Coricelli G, Critchley HD, Joffily M, O’Doherty JP, Sirigu A,
Dolan RJ: Regret and its avoidance: a neuroimaging study of
choice behaviour. Nat Neurosci 2005, 8:1255-1262.
122. De Martino B, Kumaran D, Seymour B, Dolan RJ: Frames, biases,
and rational decision-making in the human brain. Science
2006, 313:684-687.
123. Fehr E, Camerer CF: Social neuroeconomics: the neural circuitry
of social preferences. Trends Cogn Sci 2007, 11:419-427.
Current Opinion in Neurobiology 2008, 18:185–196
196 Cognitive neuroscience
124. Chiu PH, Kayali MA, Kishida KT, Tomlin D, Klinger LG, Klinger MR,
Montague PR: Self responses along cingulate cortex reveal
quantitative neural phenotype for high-functioning autism.
Neuron 2008, 57:463-473.
125. Knoch D, Pascual-Leone A, Meyer K, Treyer V, Fehr E:
Diminishing reciprocal fairness by disrupting the right
prefrontal cortex. Science 2006, 314:829-832.
126. Shoham Y, Powers R, Grenager T: If multi-agent learning is the
answer, what is the question? Artif Intell 2007, 171:365-377.
127. Gmytrasiewicz P, Doshi P: A framework for sequential planning
in multi-agent settings. J Artif Intell Res 2005, 24:49-79.
128. Breland K, Breland M: The misbehavior of organisms. Am
Psychol 1961, 16:681-684.
129. Talmi D, Seymour B, Dayan P, Dolan RJ: Human Pavlovian-
instrumental transfer. J Neurosci 2008, 28:360-368.
130. El-Amamy H, Holland PC: Dissociable effects of disconnecting
amygdala central nucleus from the ventral tegmental area or
substantia nigra on learned orienting and incentive
motivation. Eur J Neurosci 2007, 25:1557-1567.
131. Murschall A, Hauber W: Inactivation of the ventral tegmental
area abolished the general excitatory influence of Pavlovian
cues on instrumental performance. Learn Mem 2006, 13:123-
126.
132. Graeff FG, Guimares FS, Andrade TGD, Deakin JF: Role of 5-HT in
stress, anxiety, and depression. Pharmacol Biochem Behav
1996, 54:129-141.
133. Soubrié P: Reconciling the role of central serotonin neurons in
human and animal behaviour. Behav Brain Sci 1986, 9:364.
134. Dayan P, Huys QJM: Serotonin, inhibition, and negative mood.
PLoS Comput Biol 2008, 4:e4.
135. Blanchard DC, Blanchard RJ: Ethoexperimental approaches to
the biology of emotion. Annu Rev Psychol 1988, 39:43-68.
136. Mobbs D, Petrovic P, Marchant JL, Hassabis D, Weiskopf N,
Seymour B, Dolan RJ, Frith CD: When fear is near: threat
imminence elicits prefrontal-periaqueductal gray shifts in
humans. Science 2007, 317:1079-1083.
137. Dayan P, Niv Y, Seymour B, Daw ND: The misbehavior of value
and the discipline of the will. Neural Netw 2006, 19:1153-1160.
138 . Frank MJ, Samanta J, Moustafa AA, Sherman SJ: Hold your
 773.
horses: impulsivity, deep brain stimulation, and medication in
Parkinsonism. Science 2007, 318:1309-1312 An earlier paper by
the same group [141] suggested that the subthalamic nucleus
(STN) might play a crucial role in slowing down behavior when a
choice has to be made between strongly appetitive options, so the
absolute best one can be chosen without impulsive early choices.
This very clever study uses two different treatments for
Parkinson’s disease (dopaminergic therapy and deep brain
stimulation of the STN) together with a very important prediction/
action learning task [86] to test this hypothesis, and the distinction
between STN and dopaminergic effects in the striatum..
139. Aron AR, Poldrack RA: Cortical and subcortical contributions to
stop signal response inhibition: role of the subthalamic
nucleus. J Neurosci 2006, 26:2424-2433.
140. Aron AR, Behrens TE, Smith S, Frank MJ, Poldrack RA:
Triangulating a cognitive control network using diffusion-
weighted magnetic resonance imaging (MRI) and functional
MRI. J Neurosci 2007, 27:3743-3752.
141. Frank MJ: Hold your horses: a dynamic computational role for
the subthalamic nucleus in decision making. Neural Netw 2006,
19:1120-1136.
Current Opinion in Neurobiology 2008, 18:185–196
142. Atallah HE, Lopez-Paniagua D, Rudy JW, O’Reilly RC: Separate
neural substrates for skill learning and performance
in the ventral and dorsal striatum. Nat Neurosci 2007,
10:126-131.
143. Belin D, Everitt BJ: Cocaine seeking habits depend upon
dopamine-dependent serial connectivity linking the ventral
with the dorsal striatum. Neuron 2008, 57:432-441.
144. Joel D, Weiner I: The connections of the dopaminergic system
with the striatum in rats and primates: an analysis with respect
to the functional and compartmental organization of the
striatum. Neuroscience 2000, 96:451-474.
145. Haber SN, Fudge JL, McFarland NR: Striatonigrostriatal
pathways in primates form an ascending spiral from the shell
to the dorsolateral striatum. J Neurosci 2000, 20:2369-2382.
146. Barto A, Mahadevan S: Recent advances in hierarchical
reinforcement learning. Discrete Event Dynamic Systems 2003,
13:341-379.
147. Botvinick MM: Hierarchical models of behavior and prefrontal
function. Trends Cogn Sci 2008, 12:201-208.
148. Bouton ME: Context, ambiguity, and unlearning: sources of
relapse after behavioral extinction. Biol Psychiatry 2002,
60:322-328.
149. Redish AD, Jensen S, Johnson A, Kurth-Nelson Z: Reconciling
reinforcement learning models with behavioral extinction and
renewal: implications for addiction, relapse, and problem
gambling. Psychol Rev 2007, 114:784-805.
150. Montague PR, McClure SM, Baldwin PR, Phillips PEM,
Budygin EA, Stuber GD, Kilpatrick MR, Wightman RM: Dynamic
gain control of dopamine delivery in freely moving animals. J
Neurosci 2004, 24:1754-1759.
151. Tobler PN, Dickinson A, Schultz W: Coding of predicted reward
omission by dopamine neurons in a conditioned inhibition
paradigm. J Neurosci 2003, 23:10402-10410.
152. Tobler PN, Fiorillo CD, Schultz W: Adaptive coding of reward
value by dopamine neurons. Science 2005, 307:1642-1645.
153. Preuschoff K, Bossaerts P: Adding prediction risk to the theory
of reward learning. Ann NY Acad Sci 2007, 1104:135-146.
154. Lammel S, Hetzel A, Hckel O, Jones I, Liss B, Roeper J: Unique
properties of mesoprefrontal neurons within a dual
mesocorticolimbic dopamine system. Neuron 2008, 57:760-
155. O’Reilly RC, Frank MJ: Making working memory work: a
computational model of learning in the prefrontal cortex and
basal ganglia. Neural Comput 2006, 18:283-328.
156. Schultz W, Dayan P, Montague PR: A neural substrate of
prediction and reward. Science 1997, 275:1593-1599.
157. Gibbon J, Malapani C, Dale C, Gallistel C: Toward a neurobiology
of temporal cognition: advances and challenges. Curr Opin
Neurobiol 1997, 7:170-184.
158. Day JJ, Roitman MF, Wightman RM, Carelli RM: Associative
learning mediates dynamic shifts in dopamine signaling in the
nucleus accumbens. Nat Neurosci 2007, 10:1020-1028.
159. Hyland BI, Reynolds JNJ, Hay J, Perk CG, Miller R: Firing modes
of midbrain dopamine cells in the freely moving rat.
Neuroscience 2002, 114:475-492.
160. Pan WX, Schmidt R, Wickens JR, Hyland BI: Dopamine cells
respond to predicted events during classical conditioning:
evidence for eligibility traces in the reward-learning network. J
Neurosci 2005, 25:6235-6242.
www.sciencedirect.com